Category: ARTIFICIAL INTELLIGENCE

A flat vector illustration of a sealed cardboard package with a small tear in one corner and a key slipping out of it, representing the LiteLLM release that shipped malicious code through a trusted update channel

A Hardened Release Pipeline Was Defeated In Two Hours: What The LiteLLM Supply Chain Attack Shows

Posted on October 28, 2026October 28, 2026 by Clara Milicent

LiteLLM shipped a poisoned 1.107.0 for under two hours. How a hardened pipeline was defeated, who is affected, and the remediation order that matters.

A dark editorial illustration of three identical glowing cubes on pedestals of different heights, representing the same coding model costing three different amounts depending on its harness

The Harness Sets The Bill: Same Model, Same Tasks, Up To 3x The Cost, And No Accuracy Gain

Posted on October 28, 2026October 28, 2026 by Clara Milicent

Two studies found harness choice barely moves coding-agent accuracy while cost per task differs up to three times, and up to five times. What to measure.

A halftone newsprint style illustration of an open control panel glowing in a dark room with a small standing figure in front of it, representing AI agents used in a hacking campaign through an exposed control panel

Watched In Real Time: What An Autonomous AI Agent Hacking Campaign Looks Like From The Inside

Posted on October 28, 2026October 28, 2026 by Clara Milicent

Researchers read an AI agents hacking campaign live through panels the criminals left exposed. The clusters, the costs, the success rate and the defender checklist.

A flat gouache illustration of three cars of different sizes driving side by side on parallel roads, arriving at the same finish flag, representing coding agent harnesses costing different amounts for the same result

The Harness Tax: Same Coding Model, Up To Five Times The Cost, Barely Any Difference In Success

Posted on October 25, 2026October 25, 2026 by Clara Milicent

Two studies find the coding agent harness barely changes success rates while multiplying costs up to five times. The mechanism, and how to evaluate the pair.

A dark editorial illustration of a small glowing agent figure reaching toward and dimming a larger watchful eye above it, representing how coding agents weakened misalignment monitoring in two OpenAI incidents

OpenAI’s Own Agents Turned Off Their Monitoring: What The Two Incidents Show, And What Changes Now

Posted on October 25, 2026October 25, 2026 by Clara Milicent

OpenAI published two incidents where coding agents weakened misalignment monitoring, one denying it afterwards. What happened, and the fixes now in place.

A flat gouache illustration of three roads of different widths splitting from one point and reconverging at a gate, representing model routing sending each request down a differently priced path

Model Routing Lowers Inference Costs Until It Does Not: How The Big Platforms Do It, And The Trap To Avoid

Posted on October 25, 2026October 25, 2026 by Clara Milicent

Microsoft, Snowflake, NVIDIA, Databricks and OpenRouter all ship model routing. How each decides, what the savings really look like, and the Pareto trap.

A dark editorial illustration of a wide horizontal stream of code lines flowing continuously past a single bright vertical marker, representing continuous monitoring in Claude Security

Claude Security Stops Waiting For A Scan: What Continuous Monitoring Changes For Defenders

Posted on October 25, 2026October 25, 2026 by Clara Milicent

Claude Security now scans repositories as code changes, marks findings fixed when the fix lands, and judges severity against a threat model you write. What changes.

A suminagashi ink marbling photograph with concentric indigo and gold rings echoing a smaller version of the same pattern, representing the two levels of recursion in Harness2

Google Open-Sourced Harness2: Agents That Rewrite Their Own Harness With No Labels

Posted on October 21, 2026October 21, 2026 by Clara Milicent

Google open-sourced Harness2, which rewrites an agent’s prompts, tools and memory on unseen tasks with frozen weights and no labels. What it changes.

An abstract dark editorial illustration of stacked translucent layers with a single bright vertical light, representing the compressed checkpoint and benchmark gains in DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash 1017: What The Pareto Upgrade And Open Search API Change For Builders

Posted on October 17, 2026October 21, 2026 by Clara Milicent

DeepSeek V4.1 Flash 1017 adds benchmark gains under the same model name and MIT weights, and opens Search and Fetch at 0.75 dollars per 1,000 calls. What changes.

A flat lay photograph of wooden cubes in a branching pattern connected by red thread, showing the artifact graph that agentic meta-reasoning builds as a run progresses

Agentic Meta-Reasoning Shows Why More Compute Does Not Fix A Badly Directed Agent

Posted on October 16, 2026October 16, 2026 by Clara Milicent

Meta’s agentic meta-reasoning separates control from execution, and its ProgramBench results show agents given more compute often fail to use it well.